# AI in Healthcare: Where It Works, Where It Fails and Why

> AI in healthcare, with the evidence: where it works, from retinal screening to AI scribes, why sepsis and bias failures happened, and what the field demands.

- **Author:** Althaf Ashraf — AI Systems Engineer, Tata Consultancy Services (https://www.1stepgrow.com/authors/althaf-ashraf/)
- **Published:** Aug 9, 2026 · **Updated:** Sep 17, 2026
- **Topic:** Artificial Intelligence · **Format:** Guide · **Read time:** 9 min
- **Canonical URL:** https://www.1stepgrow.com/articles/ai-in-healthcare/

## Key takeaways

- The wins are concentrated in narrow, well-specified tasks with clear ground truth — imaging, triage, documentation.
- The FDA said in January 2025 that it had authorised more than 1,000 AI-enabled devices, and radiology dominates its public list.
- Most failures are deployment failures, not modelling failures: distribution shift, workflow mismatch and no clinician buy-in.
- Domain knowledge is worth more in healthcare AI than in almost any other vertical.
- Regulatory and validation literacy is a genuine differentiator for anyone entering this field.

AI in healthcare works best on narrow, well-defined tasks with clear ground truth, such as screening retinal images or drafting clinical notes. It fails most often after deployment, not in testing: when patients differ from the training data, when tools do not fit clinical workflow, or when a convenient proxy hides bias.

The gap between the two is rarely the model. When one US health system independently validated Epic's proprietary sepsis model, it missed 67% of patients with sepsis, well short of the developer's reported performance. What separated failures like that from the successes was everything that happened after the paper.

This guide is for engineers, analysts and clinicians weighing work in the field. It sets out the documented wins and the documented failures, each with its source, and then the skills that decide which side of that line a project lands on.

## Where does AI in healthcare genuinely work?

The successes share a pattern: a narrow task, clear ground truth and a human in the loop.

### Imaging and regulated devices

**Medical imaging.** This is the clearest success. In 2018 the FDA authorised IDx-DR, which screens retinal photographs for diabetic retinopathy. Its [pivotal trial](https://pmc.ncbi.nlm.nih.gov/articles/PMC6550188/) enrolled 900 participants across 10 primary care sites. It reported 87.2% sensitivity and 90.7% specificity. The authors also describe it as the first autonomous diagnostic AI system authorised by the FDA in any field of medicine. Flagging findings on radiographs and screening pathology slides follow the same template: a well-specified task, abundant labelled data and unambiguous ground truth.

**Regulated devices at scale.** Imaging is no longer a handful of pilots. By January 2025 the FDA said it had [authorised more than 1,000 AI-enabled devices](https://www.fda.gov/news-events/press-announcements/fda-issues-comprehensive-draft-guidance-developers-artificial-intelligence-enabled-medical-devices), and radiology accounts for most entries on its public [list of AI-enabled medical devices](https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices).

### Prediction, documentation and admin

**Deterioration prediction.** Models that watch vitals and lab results can flag patients at risk of sepsis or decline. They add value where they are integrated properly. They also show why alert design matters: a model that fires too many alerts gets ignored, which can make it worse than nothing.

**Clinical documentation.** Ambient AI scribes draft notes from a recorded consultation, and the clinician reviews them before they count. The evidence is promising but more modest than vendor claims. In a [randomised trial of 238 outpatient physicians](https://pmc.ncbi.nlm.nih.gov/articles/PMC12768499/) published in NEJM AI, one scribe (Nabla) cut time spent writing notes by 9.5% compared with usual care, while the other (Microsoft DAX Copilot) showed no significant change. Both tools showed possible improvements in burnout measures that still need confirming in larger trials. Clinicians also noted occasional clinically significant inaccuracies, so the review step is not optional.

**Administrative work.** Coding, scheduling, prior authorisation and claims are unglamorous, but the return on investment is easy to measure and the clinical risk is comparatively low.

| Use case | Real example | What the evidence shows | Main risk |
|---|---|---|---|
| Retinal screening | IDx-DR (2018) | 87.2% sensitivity, 90.7% specificity in a pivotal trial | Ungradable images, access to follow-up care |
| Radiology support | Devices on the FDA list | Largest category of FDA-authorised AI devices | Performance drift across scanners and sites |
| Sepsis prediction | Epic Sepsis Model | Poor discrimination in external validation | Missed cases and alert fatigue |
| Ambient documentation | DAX Copilot, Nabla | 9.5% less note time for one tool, no significant change for the other | Occasional inaccuracies in drafted notes |


## Why do healthcare AI projects keep failing?

Most failures are deployment failures. The model looked good on paper, but the real world did not match the paper.

### When the data or the setting changes

**Distribution shift.** Distribution shift means the data a model meets in use differs from the data it learned from, and it is one of the most common failure modes. A model trained at one hospital meets a different population, different scanners and different labelling conventions. It then degrades quietly. A 2018 [PLOS Medicine study of pneumonia detection on chest X-rays](https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002683) found that models performed better internally than at external hospitals in 3 of 5 comparisons. The network could also identify which hospital system an image came from with near-perfect accuracy. It was learning the site, not just the disease.

**Validation that does not survive contact.** The best-known sepsis example is Epic's proprietary sepsis model. An [external validation in JAMA Internal Medicine](https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307) covering 38,455 hospitalisations at Michigan Medicine found an area under the curve of 0.63, well below the developer's reported figures. The model missed 67% of patients with sepsis, yet it still generated alerts in 18% of all hospitalisations.

### When people and proxies are ignored

**Workflow mismatch.** A tool that is accurate but adds three clicks to an already compressed consultation will not be used. Clinical workflow is not a deployment detail; it decides whether the model has any effect at all.

**Label quality and proxy bias.** Healthcare labels are often proxies. "Diagnosed with X" means someone was coded with X, which reflects who got tested, who had access, and how the coding was done. The landmark example is a widely used care-management algorithm studied in [Science in 2019](https://pubmed.ncbi.nlm.nih.gov/31649194/). It predicted healthcare costs rather than illness, and because less money was spent on Black patients with the same needs, it under-referred them. Fixing the bias would have raised the share of Black patients receiving additional help from 17.7% to 46.5%.

**No clinician ownership.** Projects built by data teams and presented to clinicians at the end have a poor record. The ones that work involve clinicians in choosing the problem, not just validating the answer.

**Equity failures.** Models trained on populations that under-represent certain groups can perform worse for those groups. In healthcare that is not a metrics footnote. It is a harm, and increasingly a regulatory matter too.


## What does working in healthcare AI require?

Healthcare AI rewards domain fluency, validation literacy and patience more than raw modelling skill. Four things matter more here than in most other areas of applied AI.

- **Domain fluency.** Understanding why a lab value is missing — often the test was not ordered, which itself carries information — is the kind of knowledge that separates a useful model from a dangerous one. It takes deliberate effort, but it compounds.
- **Validation literacy.** Know why a strong test-set score is not clinical validation, what external and prospective validation involve, and when a tool becomes a regulated device. The sepsis and pneumonia studies above are the reason this matters, and understanding them puts you ahead of most candidates entering this space.
- **Patience with long cycles.** Procurement is slow, validation is slow, and integration with hospital systems is slow. If you need a fast feedback loop, this vertical will frustrate you.
- **A long-term view.** The trade-off is durability. Generic ML skill is increasingly commoditised, whereas clinical domain expertise combined with ML is not, and it is hard to displace.

Who should not choose this field? Anyone who wants to ship weekly, or who is uncomfortable being accountable for errors that affect patients rather than dashboards.

## Is AI in healthcare living up to the promise?

Partly. AI in healthcare is neither the revolution the 2018 headlines promised nor the disappointment the backlash suggested. As of September 2026, it is a field where narrow, carefully deployed tools deliver real value, while ambitious general-purpose systems keep meeting the same wall.

For anyone entering, the lesson is consistent. Pick a narrow problem, work with the clinicians who have it, and treat deployment as the hard part rather than the afterthought. That is where the successes have consistently come from.


## Where to go next

To test your interest, read the Epic sepsis validation linked above in full; it is the clearest short lesson in how clinical AI goes wrong. For the technical path, see our [AI engineer roadmap](https://www.1stepgrow.com/articles/ai-engineer-roadmap-2026). For how the field's terms relate, read [AI vs machine learning vs deep learning](https://www.1stepgrow.com/articles/ai-vs-machine-learning-vs-deep-learning). For the forecasting side of clinical data, [predictive analytics in healthcare](https://www.1stepgrow.com/articles/predictive-analytics-in-healthcare) goes deeper, and [what data science is and why it matters](https://www.1stepgrow.com/articles/what-is-data-science) covers the underlying discipline.

## Frequently asked questions

### What is AI actually good at in healthcare today?

Narrow, well-defined tasks with reliable ground truth: detecting specific findings in medical images, flagging deterioration risk from vitals, drafting clinical documentation for a clinician to review, and administrative work like coding and scheduling. Broad diagnostic reasoning remains far harder than headlines suggest, and even documentation tools show modest, uneven gains in trials.

### Why do so many healthcare AI projects fail?

Rarely because the model was inaccurate in testing. Usually the deployment population differed from the training population, the tool did not fit clinical workflow, or clinicians were brought in too late to trust or use it. Epic's sepsis model is the standard example: external validation found it missed most sepsis cases while still alerting on a large share of admissions.

### Do I need a medical background to work in healthcare AI?

No, but you need to build domain fluency deliberately and work closely with clinicians. Teams that treat healthcare as just another tabular dataset produce models that are technically sound and clinically useless. A practical start is learning why values go missing, how diagnoses get coded, and which decision a clinician actually makes at the point your model would fire.

### Is healthcare AI a good career area?

Yes, if you can accept slow cycles. Procurement, validation and hospital integration all take time, and regulatory overhead is real. In exchange, clinical domain expertise compounds unusually well and is much harder to displace than generic ML skill. It suits people who prefer durable, accountable work over shipping something new every week.

### What regulatory knowledge is needed?

Enough to know when a tool becomes a regulated medical device, what clinical validation requires beyond a test-set score, and how data governance and patient privacy constrain what you can build. In practice that means understanding external and prospective validation, not just held-out accuracy. Few candidates arrive with this literacy, so it is a strong differentiator.

---
_Source: 1stepGrow (https://www.1stepgrow.com/articles/ai-in-healthcare/). Cite with the title, "1stepGrow" and a link._